The main goal of the thesis is to develop an automated anomaly detection system for vessel data to enhance monitoring and reporting of vessels' energy management.

The thesis is driven by existing research in the field to find common anomalies in maritime sensors and get motivation to select an appropriate machine learning model that can be used as an anomaly detector without much human intervention.

To achieve the aim of an autonomous machine learning-based anomaly detector for vessel data, the data is pre-processed with feature selection, and popular unsupervised anomaly detection algorithms like Isolation Forest (iForest), k-nearest neighbours (KNN), and the Local Outlier Factor (LOF) are tested with a dataset with synthetic anomalies. The models are trained with clean data (data without anomalies) and tested with synthetic anomalies closer to real-world anomalies. Different types of anomalies are manipulated for testing purposes. A linear regression model has been tried and has exhibited overfitting phenomena.

The acceptable performances needed for anomaly detection are achieved by eXtreme Gradient Boosting (XGBoost) with proper thresholding for the residuals from the prediction. Thus, a model for an autonomous anomaly detection system is proposed.